Fast RT-LoG operator for scene text detection

Abstract

This paper proposes a new real-time Laplacian of Gaussian (RT-LoG) operator for scene text detection. This method takes advantage of the Gaussian kernel distribution in the spatial/scale-space domains and kernel decomposition with the box filtering method. Two levels of optimization are given. The first level of optimization within the spatial domain is obtained by box mutualization. The second level of optimization within the spatial/scale-space domains is performed using a mixed method for box selection. The proposed RT-LoG operator is evaluated on the ICDAR2017 RRC-MLT dataset in terms of robustness and time processing. The results are compared with the state-of-the-art real-time operators for scene text detection. The proposed operator appears as the top performance with the best trade-off between robustness and time processing. The proposed operator can support approximately 30 frames per second (FPS) up to the Quad-HD resolution on a regular CPU architecture with a low-level latency. In addition, the proposed operator can support the full pipeline for scene text detection. Our system is competitive with the top accurate systems of the literature while processing with a difference of two orders of magnitude in term of processing resources.